← Back to work

Case study · AI agent product

Designing the human-in-the-loop layer of an AI agent product.

Dubot gives businesses an admin panel to control what their AI agent can do and what context it has. I design the whole product as its solo designer. This case study covers one part of it: the end-user experience.

  • Solo product designer
  • 2 engineers, no PM or researcher
  • 2026–present
  • Live
  • dubot.ai ↗
The challenge

AI agents now take real actions in a product. End users need to approve them without it becoming a chore.

What I did

Discovery interviews, then the full end-user surface: questions, confirmation, live progress and failure states.

The result

Live, and a key part of enterprise pilot demos. Shaped with an enterprise pilot customer.

Live, interactive prototype Pick a journey from the top bar

The scope

A full surface around the agent.

  • Side panel

    I tested it against a wizard and a command bar. It won in early interviews: the agent stays present without taking over the page.

  • Status bar

    Shows what the agent is doing, with a way to pause, even when the panel is closed.

  • A second AI on the page

    When another assistant is already in the product, users always know which one is acting.

The research

Most interviewees asked for a confirmation step before seeing any design.

Dozens of interviews with prospective and pilot customers. Most asked, unprompted, whether end users would see the result first.

A secondary signal

Another company's CX team designed their own agent confirmation from scratch and landed on almost the same shape.

The pattern

Ask, gather, and always show the full payload before it fires.

Questions use real controls, like choices and short forms. Before any action, the user sees exactly what will happen and can edit or cancel it.

Figma design of a 'How can we help?' assistant panel: a getting-started guide, help center, product walkthrough, talk to support, and quick-action chips.
The entry point: a docked panel inside the product.
Figma design of the agent disambiguating which record to edit, then offering a list of selectable next actions, instead of expecting a typed chat answer.
The agent narrows down the record, then offers choices to select.
Figma design of a live step tracker showing the agent's progress toward an action, alongside a confirmation card with the exact discount and new total, plus Apply and Cancel buttons.
Live progress, then the full payload before the agent acts.

Beyond the happy path

Every failure has its own recovery.

  • Declines when unsure
  • Needs clarification
  • Blocked by an admin
  • Cancelled
  • Failed, with retry
  • Walkthrough interrupted
Figma design of a failed action state: the agent's step tracker shows Confirm and apply failed with a connection issue, alongside a retry or not-now choice.
A failed action: what went wrong and a way to retry.

Many of these came from working with an enterprise pilot customer.

What I learned

With an agent, the user journey stops being a straight line.

An agent can end up almost anywhere depending on what someone asks, so there's no fixed journey to design against. I test directions in working prototypes, so people react to behavior and not only to static screens.

Hiring for a product design role? I'd love to hear about it.